When Agents Became the Customer: Cloudflare's cf CLI Covers 3,000 API Operations and Retires Wrangler

Every generation of developer tooling eventually faces the same question: who is actually holding the keyboard? For most of computing history the answer was a person. Cloudflare’s launch of its new cf command-line tool on September 28, 2026 — announced in a Birthday Week blog post by Matt “TK” Taylor and Samuel Macleod — is built around a different assumption: the primary user of infrastructure CLIs is increasingly not a person at all, but an AI agent acting on one’s behalf.

The numbers the company disclosed are the real story. In March 2026, agents accounted for a quarter of all Wrangler usage, up from single-digit percentages a year earlier. By the last week of September, that share had climbed to 48%. Agents are not just numerous — they are power users. Cloudflare’s telemetry shows agents invoke nearly twice as many distinct commands per day as humans, and are almost four times as likely to chain six or more commands in a session. When nearly half of your CLI traffic is non-human, the design brief for the next-generation tool writes itself.

Why Wrangler had to go

Wrangler, Cloudflare’s long-serving CLI, exposes roughly 280 operations — a hand-built surface assembled product team by product team over years, with each team taking its own approach to command design. The blog post is candid about the consequences: inconsistent terminology across d1 info, hyperdrive get, and workflows describe, custom experiences spanning thousands of lines of rarely used code, and divergent solutions to identical problems. Enforcing patterns across teams was, in the company’s words, virtually impossible.

Meanwhile, the Cloudflare API itself offers thousands of operations that the CLI never surfaced. The gap between what agents could do on the platform and what they could reach through the official tool kept widening.

The fix is cf, and its most striking statistic is coverage: built on Forge, Cloudflare’s new unified API generation pipeline, cf generates its commands directly from the OpenAPI schemas that power the company’s API documentation and SDK generation. Annotate the schema, and the CLI follows. That pipeline takes cf from Wrangler’s ~280 commands to the entirety of the Cloudflare API surface — more than 3,000 operations. A single agent session can now provision a Worker, deploy it, wire up observability, protect it with Cloudflare Access, register a domain, and place the WAF in front of it, without ever leaving one tool.

Designed for a user that reads JSON, not tables

The design decisions read like a checklist of agent ergonomics. JSON output is the default interface — pretty-printed for humans, condensed for agents to save context tokens. This inverts Wrangler’s stance, where --json was an opt-in flag on only some commands and many commands returned Unicode tables designed for terminal-reading humans. Agents can filter results themselves and hand the human a formatted summary; the human is, as the post puts it, one step removed from the CLI.

Command discovery is solved with cf cli search, which lets an agent describe what it wants to do in natural language and receive a ranked list of matching commands drawn from a compact search index over API descriptions and parameters. With 3,000 routes through the tool, brute-forcing --help is a context-budget disaster; targeted retrieval is the only sane default. Cloudflare says agents are automatically told about the search command the first time they invoke --help.

Configuration moves from TOML and JSONC to cloudflare.config.ts — TypeScript, with all the type-safety that implies. Typed configuration turned out to be an agent accelerant even with zero prior exposure to the format: agents edited configurations correctly on demand, including the env semantics that changed dramatically from Wrangler. Editors and agent harnesses that speak the Language Server Protocol — Claude Code and Codex are named explicitly — get autocomplete and inline documentation through the same plugin machinery. Internally, Cloudflare reports some configuration files shrank by 40% from more than 5,000 lines by defining environments programmatically from a shared base instead of copy-pasting env blocks.

Finally, cf is built on Vite by default — the Cloudflare Vite Plugin becomes the recommended build path for Workers, with hot module replacement in local development and Rolldown-based tree-shaking for builds. Wrangler’s :8787 dev server and its esbuild bundling remain supported through delegation, but they are legacy paths now.

A new tool rather than a new version

The most counterintuitive decision is shipping a brand-new binary instead of updating Wrangler. Cloudflare’s reasoning is grounded in how LLMs actually behave: years of Wrangler documentation, blog posts, and third-party guides are baked into model training data. Radically changing a familiar tool fights that learned behavior — the model keeps reaching for old flags and old semantics. A tool the model has never seen, introduced with explicit context injections and appended AGENTS.md guidance, turns out to be less confusing than teaching an agent to unlearn. The post includes a copy-paste prompt agents can use to internalize the differences.

Migration tooling acknowledges reality rather than pretending it away. cf migrate converts Vite-built Workers to the new configuration format automatically, and Wrangler-built projects continue to work — cf delegates builds to Wrangler where esbuild is still required. When the open beta ends, Cloudflare will ship a final major Wrangler release that points users and agents to cf, followed by 18 months of maintenance support. The tool that defined the Workers ecosystem for a decade gets a defined runway, not an abrupt end.

The signal for everyone else

Two broader points make this launch worth watching beyond the Cloudflare ecosystem.

First, it is a concrete, telemetry-backed data point for the agent-era economy. A year of internal numbers — single digits to 25% to 48% in about twelve months — is the steepest non-human adoption curve a mainstream developer tool has published. When Cloudflare’s founders wrote their Annual Founders’ Letter two days earlier warning that automated traffic is on its way to 1,000x human volume, this is the developer-facing mirror of the same shift: agents are not coming for developer tools, they are already the median user.

Second, it establishes a template for agent-first interface design that other platform vendors will likely copy: machine-readable output by default, retrieval-based command discovery instead of memorization, strongly typed configuration that language servers can explain, and system-prompt-friendly documentation. The post itself predicts these patterns “will become standard in more CLIs in the near future.”

cf is available in open beta via npm i -g cf, is open source with issues tracked on GitHub, and works from any directory. For teams whose deploy pipelines assume Wrangler, the 18-month clock has started. For everyone else building platforms that agents are expected to operate, September 28 may be remembered as the day a major infrastructure company officially redesigned its developer surface around a non-human customer.